CROP-TYPE CLASSIFICATION BASED ON TIME-SERIES OF SPECTRAL INDICES AND MACHINE LEARNING FOR HYDROLOGICAL STUDIES

Watershed-scale studies of non-point source pollutants in agricultural systems depend on yearly detailed crop information to characterize in-field management practices. In the U.S., yearly crop data layers (CDL) are available starting in the early 2000s from the U.S. Department of Agriculture. These datasets were developed using combined field information and sets of temporal imagery from multiple sensors. Development of such data layers in the 1980s and 1990s is deterred by the lack of field information and the reduced number of satellite-based multi-spectral imagery. In this study, a procedure to estimate historical crop type layers is proposed and evaluated. Time series of NDVI datasets from Landsat 5 TM sensor for two watersheds located in North Dakota and Mississippi for years 1985, 1990, 1995, 2000, and 2005 were collected and processed. Pixel-based information was aggregated into field-scale to minimize the effect of mixed pixels. The Weighted Window Linear Regression algorithm was used to remove noise such as clouds, shadows, snow, data gaps, and water in the phenology curves. Reference phenology curves for six crop types were developed based existing CDL datasets from 2000 and 2005. Models using decision trees, ANN, logistic model trees, and random forest algorithms were evaluated using the 2000 and 2005 datasets. Random forest algorithm was selected. Input datasets for all years were post processed using non-linear regression to address differences in input dimensionality given by the yearly differences in the number of images and their respective acquisition dates. Classifiers were generated using the 2000 and 2005 datasets and evaluated using CDL layers as reference for quantitative accuracy assessment. Results indicate sensitivity of models to yearly variations in phenology curves and site-specific characteristics. Conversely, the models developed using the 2000 imagery demonstrated spatial-temporal generalization potential (Kappa of 0.65). Additionally, historic estimates for years 1985, 1990, and 1995 based on the 2000 model were compared to county-scale crop records and matched long-term trends for individual crops, signifying the potential of estimated pseudo-CDL layers to support the long-term input generation for hydrological models.